Abstract: To address the issues of numerous small targets, significant scale variations, and complex background interference that limit the detection accuracy in aerial image small target detection, a small target detection algorithm for Unmanned Aerial Vehicle (UAV) aerial images, named CSL-YOLOv12, was proposed based on the improved YOLOv12n. Firstly, a CSP-Partial Convolution (CSP-PConv) module was designed to improve the backbone network, which enhanced the model's ability to extract context information while reducing the computational load. Secondly, a Self Adaptive Calibration Feature Pyramid Network (SACFPN) was designed to improve the neck network, which enhanced the model's multi-scale feature fusion capability. Finally, a Lightweight Shared Convolutional detection Head (LSCHead) was introduced to improve the model's localization and classification of small targets. Experimental shows that the improved algorithm achieves a mean Average Precision (mAP) and accuracy of 39.6% and 49.2% on the VisDrone2019 dataset, respectively, which has improved 5 percentage points and 3.1 percentage points respectively than those of the baseline model YOLOv12n. The recall rate has increased by 4 percentage points to 38.2%. The improved algorithm effectively enhances the detection accuracy in aerial scenes.

Key words: Unmanned Aerial Vehicle (UAV), small target detection, multi-scale feature fusion, feature pyramid, shared convolutional

摘要： 针对航拍图像小目标检测中存在的小目标数量多、尺度变化大、复杂背景干扰导致的检测精度受限等问题，提出一种基于YOLOv12n改进的无人机航拍图像小目标检测算法CSL-YOLOv12。首先，设计部分多尺度特征提取模块(CSP-PConv)模块改进主干网络，增强模型上下文信息提取能力,同时减少模型计算量；其次，设计一种自适应校准特征金字塔结构(SACFPN)，改进颈部网络，提升模型多尺度特征融合能力；最后，引入轻量共享卷积检测头(LSCHead)，增强模型对小目标的定位与分类。实验结果表明，改进算法在VisDrone2019数据集上的平均精度均值(mAP)与准确率达到39.6%和49.2%，相较于基准模型YOLOv12n分别提升了5个百分点和3.1个百分点，召回率提升了4个百分点达到38.2%。验证了改进算法有效提升了航拍场景下的目标检测精度。

关键词: 无人机, 小目标检测, 多尺度特征融合, 特征金字塔, 共享卷积